Chinese artificial intelligence developers have made notable progress training AI models using domestic chips, particularly in inference and post-training phases, even as crucial pre-training on local hardware remains limited compared to Nvidia-powered US models.
- Domestic chips widely used for inference, limited for pre-training
- Zhipu AI and Meituan lead with models trained on Huawei accelerators
- Startups prove efficiency of lightweight models on local hardware
What happened
Chinese AI researchers and companies have increasingly utilized domestic silicon chips for training and running AI models amid US export controls and Beijing’s technological self-sufficiency initiatives. Models such as Zhipu AI's GLM-Image and Meituan's LongCat-2.0 have been trained or run on clusters powered entirely by Chinese hardware, primarily Huawei’s Ascend series.
Startups like ModelBest have demonstrated the viability of lightweight models optimized for smartphones and vehicles using domestic acceleration frameworks. Meanwhile, post-training of large-scale models like DeepSeek-V4-Pro has also been completed on local clusters, although the most computationally intense pre-training phase for large language models largely remains reliant on foreign silicon technologies.
Why it matters
China’s use of indigenous AI chips throughout multiple stages of development signifies a strategic effort to decrease reliance on US technology and build a comprehensive domestic AI ecosystem. This is particularly important as Washington tightens export controls on high-performance semiconductors critical to AI advancement.
Although local chips currently lag behind Nvidia-powered hardware in raw computational capacity and efficiency, sustained investment and experimentation are enabling Chinese AI labs to make meaningful progress. The development of fully domestic multimodal models and efficient on-device AI positions China for longer-term competitiveness in global AI innovation.
What to watch next
The critical next step will be pushing large-scale, compute-heavy pre-training of advanced AI models onto domestic silicon to reduce dependency on foreign accelerators. Progress from labs like Zhipu AI on migrating large language model training onto Huawei hardware will be key indicators of closing this gap.
Monitoring emerging startups and research groups developing compact, high-performance AI models optimized for native hardware will also provide insight into China’s evolving AI capabilities. Additionally, government policies and investments aimed at strengthening indigenous chip design and fabrication may accelerate the pace of technological self-reliance in AI.